Optimize I/O and add hourly LSTM support for summed Q' validation

[taddyb]

Summary

• StreamflowReader: Replace xr.interp(method="nearest") with np.repeat(..., 24) for daily→hourly interpolation, eliminating xarray interpolation overhead on every training batch
• summed_q_prime.py: Pre-filter divides to only upstream basins (mirrors training dataloader), use isel with integer indices instead of lazy sel/resample, GPU-accelerate per-gauge nansum via CuPy, and collapse hourly→daily via numpy reshape+mean on CPU before moving to GPU
• CLAUDE.md: Comprehensive rewrite with full module map, config system, training/testing pipeline, temporal resolution handling, and dataset layer documentation
• Tests: New test suite for summed Q' validating hourly/daily equivalence, pre-filtered divide correctness, store offset handling, and CuPy vs NumPy reference matching
• clean_lstm_lateral_inflows.py: Added to references for LSTM data preparation

Test plan

• uv run pytest tests/scripts/test_summed_q_prime.py — 7 tests pass
• uv run pytest tests/io/test_readers.py — 28 tests pass
• Pre-commit hooks (ruff, mypy, format) all pass
• Daily LSTM summed Q' output verified identical (float32 tolerance) to v0.5.2 baseline across 2808 gauges

🤖 Generated with Claude Code